# OpenAI's GPT-OSS-20B Model and Safety Alignment Issues in a Low-Resource Language

arXiv: https://arxiv.org/abs/2510.01266 · HTML: https://mlsanity.com/paper/2510.01266/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Safety alignment and jailbreak attacks, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 5 of its 9 references that are in the map.
- First posted on arXiv 26 Sep 2025.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Safety alignment and jailbreak attacks](https://mlsanity.com/alignment/safety-alignment-and-jailbreak-attacks/)
- LLM safety and security / [Safety alignment, refusal and guardrails](https://mlsanity.com/safety/safety-alignment-refusal-and-guardrails/)

## Builds on (most cited first, 5 of 5)

- [Reasoning Models Don't Always Say What They Think](https://mlsanity.com/paper/2505.05410/) (2025)
- [Alignment faking in large language models](https://mlsanity.com/paper/2412.14093/) (2024)
- [AI Sandbagging: Language Models can Strategically Underperform on Evaluations](https://mlsanity.com/paper/2406.07358/) (2024)
- [Large Language Models Often Know When They Are Being Evaluated](https://mlsanity.com/paper/2505.23836/) (2025)
- [Fully Autonomous AI Agents Should Not be Developed](https://mlsanity.com/paper/2502.02649/) (2025)

## Built on by (most cited first, 1 of 1)

- [BanglaMedQA and BanglaMMedBench: Evaluating Retrieval-Augmented Generation Strategies for Bangla Biomedical Question Answering](https://mlsanity.com/paper/2511.04560/) (2025)

## Abstract (arXiv)

In response to the recent safety probing for OpenAI's GPT-OSS-20b model, we present a summary of a set of vulnerabilities uncovered in the model, focusing on its performance and safety alignment in a low-resource language setting. The core motivation for our work is to question the model's reliability for users from underrepresented communities. Using Hausa, a major African language, we uncover biases, inaccuracies, and cultural insensitivities in the model's behaviour. With a minimal prompting, our red-teaming efforts reveal that the model can be induced to generate harmful, culturally insensitive, and factually inaccurate content in the language. As a form of reward hacking, we note how the model's safety protocols appear to relax when prompted with polite or grateful language, leading to outputs that could facilitate misinformation and amplify hate speech. For instance, the model operates on the false assumption that common insecticide locally known as Fiya-Fiya (Cyphermethrin) and rodenticide like Shinkafar Bera (a form of Aluminium Phosphide) are safe for human consumption. To contextualise the severity of this error and popularity of the substances, we conducted a survey (n=61) in which 98% of participants identified them as toxic. Additional failures include an inability to distinguish between raw and processed foods and the incorporation of demeaning cultural proverbs to build inaccurate arguments. We surmise that these issues manifest through a form of linguistic reward hacking, where the model prioritises fluent, plausible-sounding output in the target language over safety and truthfulness. We attribute the uncovered flaws primarily to insufficient safety tuning in low-resource linguistic contexts. By concentrating on a low-resource setting, our approach highlights a significant gap in current red-teaming effort and offer some recommendations.